risk-bound

This paper introduces PRISM, a geometric risk bound that decomposes model drift in post-training LLM variants into scale, shape, and head axes to diagnose specific failure modes like quantization errors or catastrophic forgetting.